PatchCore is an anomaly detection method that learns a compact representation of image features using a coreset sampling technique. This repository provides the implementation of PatchCore, as described in Roth et al. (2021). PatchCore extracts features from a subset of image patches via a learned feature map, enabling efficient anomaly detection. The system achieves high performance and is designed for practical application.
This project offers high accuracy in both image-level and pixel-level anomaly detection, achieving up to 99.6% AUROC for image-level and 98.4% for pixel-level. It includes pretrained models for the MVTec AD benchmark, allowing for quick evaluation and deployment. The design allows for flexible customization in terms of backbones, layers, and training parameters for adaptation to different datasets and hardware configurations.
- Image-Level Anomaly Detection: Provides accurate identification of anomalous images within a dataset.
- Pixel-Level Localization: Pinpoints the exact location of anomalies within an image.
- Pretrained Models: Includes pre-trained PatchCore models for the MVTec AD benchmark for rapid experimentation.
- Flexibility: Supports various backbones, layer configurations, and training parameters.
- Scalability: Designed to handle large datasets and complex image sizes.
- Reproducibility: Offers clear instructions and scripts for replicable training and evaluation.
- Extensibility: Supports ensembling multiple backbones for enhanced performance.
The project is mature with a well-defined implementation and detailed documentation. The repository includes extensive instructions, sample runs, and evaluation methods. Regular commits and active community support suggest ongoing maintenance and improvements. The provision of pre-trained models further indicates the project's stability and usability.
This project benefits researchers and practitioners seeking robust and accurate anomaly detection solutions. It addresses the need for efficient and effective methods to identify and locate anomalies in image data, particularly within industrial inspection settings. PatchCore provides a valuable tool for quality control, predictive maintenance, and other applications where anomaly detection is crucial.